Sift Science, Inc.

Fred Sadaghiani

Alex Paino

Jacob Burnim

Keren Gu

Noah Grant

Eugenia Ho

Gary Lee

Doug Beeferman

Micah Wylde

Wei Liu

Thomas Schiavone

Vera Dadok

Janice Lan

Aaron Beppu

Megan Mann

Aleksandr Lopatin

A machine learning-based system and method for identifying digital threats includes a threat service that: implements a unified threat model that produces a unified threat score that predicts both of: a level of threat of a piece of online content, and a level of threat that a target user will create a harmful piece of online content; wherein: implementing the unified threat model includes: receiving event data comprising historical content data for the target user and content data of the pending piece of online content and historical user digital activity data and real-time user activity data; and providing input of content feature data and user digital activity feature data to the unified threat model; and the unified threat model produces the unified threat score based on the content and the user digital activity data; and computes a threat mitigation action based on an evaluation of the threat score.

Systems and methods configuring a unified threat machine learning model for joint content and user threat detection

Systems and methods for implementing an interpretive proxy model includes evaluating a deep learning model; estimating a subset of a plurality of distinct algorithmic features of the deep learning model as leading contributors of a subject predictive output of the deep learning model; constructing a proxy model using algorithmic features of the deep learning model within the subset of the plurality of distinct algorithmic features; training the proxy model to mirror predictive outputs of the deep learning model; deploying the proxy model alongside the deep learning model based on a completion of the training; and in response to a same input to both the deep learning model and the proxy model, exposing: (1) a predictive output of the deep learning model, and (2) an explanation of the predictive output of the deep learning model based on leading contributing algorithmic features of the proxy model.

Systems and methods for configuring and implementing an interpretive surrogate machine learning model

A machine learning-based system and method for identifying digital threats that includes implementing a machine learning-based digital threat mitigation service over a distributed network of computers; constructing, by the machine learning-based digital threat mitigation service, a subscriber-specific machine learning ensemble that includes a plurality of distinct machine learning models, wherein each of the plurality of distinct machine learning models is configured to perform a distinct machine learning task for identifying a digital threat or digital fraud; constructing a corpus of subscriber-specific digital activity data for training the plurality of distinct machine learning models of the subscriber-specific ensemble; training the subscriber-specific ensemble using at least the corpus of subscriber-specific digital activity data; and deploying the subscriber-specific ensemble.

Systems and methods configuring a subscriber-specific ensemble of machine learning models

Systems and methods include: receiving digital event type data that define attributes of a digital event type; receiving digital fraud policy that defines a plurality of digital processing protocols; transmitting via a network the digital event data and the digital fraud policy to a remote digital fraud mitigation platform; using the digital event data to configure a first computing node comprising an events data application program interface or an events data computing server to detect digital events that classify as the digital event type; using digital fraud policy to configure a second computing node comprising a decisioning API or a decisioning computing server to automatically evaluate and automatically select one digital event processing outcome of a plurality of digital event processing outcomes that indicates a disposal of the digital events classified as the digital event type; and implementing a digital threat mitigation application process flow that evaluates digital event data.

System and methods for dynamic digital threat mitigation

Systems and methods include implementing a review queue interface that includes: a review queue comprising a listing of distinct review items; a current state for each of distinct review items; a listing for each review item of the distinct review items of one or more client browsers that are interacting with each review item; identifying client browser activity of the one or more client browsers; computing a computed state for each of distinct review items based on the client browser activity; computing changes to the state of review items based on an assessment of the current state and the computed state for each of distinct review items; and automatically updating a state of one or more of the distinct review items within the review queue interface based on a difference between the current state and the computed state of the one or more of the distinct review items.

Systems and methods for intelligently implementing a machine learning-based digital threat mitigation service

Systems and methods include: collecting digital threat scores of an incumbent digital threat machine learning model; identifying incumbent and successor digital threat score distributions; identifying quantiles data of the incumbent digital threat score distribution; collecting digital threat scores of a successor digital threat machine learning model; calibrating the digital threat scores of the successor digital threat score distribution based on the quantiles data of the incumbent digital threat score distribution and the incumbent digital threat score distribution; and in response to remapping the digital threat scores of the successor digital threat score distribution, publishing the successor digital scores in lieu of the incumbent digital threat scores based on requests for digital threat scores.

Systems and methods for calibrating a machine learning model

Systems and methods include: collecting digital event data for the digital account; using a trained machine learning model to extract account takeover (ATO) risk features from the collected digital event data; evaluating the extracted ATO risk features of the collected digital event data of the digital account against a plurality of ATO risk heuristics; identifying one or more of the plurality of ATO risk heuristics that is triggered by the extracted ATO risk features, wherein one or more of the plurality of ATO risk heuristics may be triggered if at least a subset of the extracted ATO risk features matches requirements of the one or more ATO risk heuristics; and generating an ATO risk assessment for the digital account using the one or more triggered ATO risk heuristics.

System and methods for digital account threat detection

Systems and methods include implementing a remote machine learning service that collects digital event data; collecting incumbent digital threat scores generated by an incumbent machine learning model and successor digital threat scores generated by a successor digital threat machine learning (ML) model; implementing anomalous-shift-detection that detects whether the successor digital threat scores of the successor digital threat ML model produces an anomalous shift; if the anomalous shift is detected by the machine learning model validation system, blocking a deployment of the successor digital threat model to a live ensemble of digital threat scoring models; or if the anomalous shift is not detected by the machine learning model validation system, deploying the successor digital threat ML model by replacing the incumbent digital threat ML model in a live ensemble of digital threat scoring models with the successor digital threat ML model.

Systems and methods detecting and mitigating anomalous shifts in a machine learning model

Systems and methods include: implementing a first machine learning model to generate an output of a global digital threat score for an online activity based on an input of the collected digital event data; implementing a second machine learning model that generates a category inference of a category of digital fraud or a category of digital abuse from a plurality of digital fraud or digital abuse categories; selecting a third machine learning model from an ensemble of digital fraud or digital abuse machine learning models based on the category inference generated by the second machine learning model, wherein the ensemble of digital fraud or digital abuse machine learning models comprise a plurality of disparate digital fraud or digital abuse category-specific machine learning models; and implementing the selected third machine learning model to generate a digital fraud or digital abuse category-specific threat score based on the digital event data.

Growing community of inventors

San Francisco, CA, United States of America

Fred Sadaghiani